Multi-class satellite imagery classification using deep learning approaches

Abstract

Recent advancements in remote sensing technologies, as well as high-resolution satellite images, have opened up new avenues for comprehending the earth's surfaces. However, owing to the significant unpredictability in satellite data, satellite images categorization is a difficult task. Availability of the satellite dataset is a challenging task in the field of remote sensing. To overcome this challenge a novel sentinel-2 image dataset is proposed. Two different techniques for categorizing a large-scale dataset containing various types of land-use and land-cover surfaces are proposed and compared for this goal. In this article, an enhanced version of ResNet50 has been proposed to predict the multiple classes from sentinel 2 images. Furthermore, the outcome of ResNet50 is compared with traditional (shallow) machine learning models and deep learning models to check the working efficiency of the proposed approach. The shallow approach had the best F1-score of 0.87, while the deep approach ResNet50 achieved the best F1-score of 0.924. It has been realized from the outcome that the deep learning approaches are most robust than the machine learning approach in terms of classifying the multi-label satellite images classification.